Artificial intelligence is having its breakout moment. Once confined to the realm of science fiction, it seems bright-eyed entrepreneurs everywhere are now getting into AI. As renowned author, editor, and futurist Kevin Kelly puts it, “The business plans of the next 10,000 startups are easy to forecast: Take X and add AI.”

But while disruptive opportunities abound for AI, there’s also no shortage of challenges to overcome. Anyone who has tried their hand at training a fledgling AI will tell you, there’s one hurdle that eclipses them all: removing a human from the loop.

It’s no secret that scaling AI typically involves human agents operating as a safety net, working in the background and ready to take the controls when a bot gets stuck. Even Facebook, an early pioneer in AI and intelligent chatbots, needs an occasional human hand to ensure a high-quality user experience.

While this offers an effective way to get AI off the ground, complete autonomy is the ultimate goal. How do you get there? As we learned firsthand building Mezi, it starts with picking the right vertical and then mastering the hand-off.

Choosing the right vertical

One of the greatest misconceptions about the field of artificial intelligence is that there will be “one AI to rule them all.” In reality, we’re far more likely to see an AI landscape dotted with countless highly specialized AIs than one dominated by artificial general intelligence (AGI).

For this reason, picking the right vertical — the right type of specialization — is one of the most important decisions to make. When we first launched Mezi, we focused on a specific experience (shopping) and tested it on a handful of verticals, such as fashion, gifting, and travel. We quickly learned that the travel industry has all of the key ingredients for training and scaling an AI.

For one, the travel industry is highly fragmented. There’s no single app that can take care of booking and coordinating travel experiences from start to finish, from booking tickets to building itineraries. The industry is also highly commoditized. It’s easy to parse the different attributes of flights, hotels, cars, tours, and the like because the metadata is already out there in a structured format that AI can easily work with.

This is the type of environment where an AI can thrive. By employing natural language processing (NLP), deep learning, and neural networks, we can more easily understand user intent, action, location, and other attributes with high degrees of accuracy. This structured environment also makes it easier to understand user preferences and run a series of machine learning algorithms to provide highly personalized recommendations.

Another useful attribute to consider when selecting a vertical for an AI is the extent of business use cases in the market. In the case of travel, a considerable portion of the market is driven by business travelers. This not only creates an opportunity for repeat use and interaction — key goals for any software application — but it also means there’s a precedent for personal assistants. Business travelers already rely upon assistants to help them with booking trips, which makes them more likely to adopt a technology solution that can help them do the job better and faster.

Mastering the hand-off

It’s also important to keep in mind that making the leap from human-supported machines to fully autonomous humanlike machines is not some momentous, “flipping the switch” event — it’s an evolution that takes place one small hand-off at a time.

When we first launched Mezi, almost every single user interaction involved a human operating behind the scenes. We hired people who were cool, calm, and experienced customer happiness specialists and put them behind the wheel. Then we started building an AI that watched and learned from their interactions with customers — their tone, word choices, emojis, everything.

Before long, we began handing off some of these tasks to the Mezi AI — at first, simple parts of the human agents’ workflow, then ratcheting up to more complex tasks from there. Each time the AI takes on a new task, the human is freed up to focus on more complex challenges and identify new areas for the AI to study.

When we launched a pilot with one of the world’s largest financial services companies in early 2017, for instance, we experienced a massive spike in traffic and our AI — which by then was doing quite well — started to stumble. Our NLP systems weren’t prepared for such diversity in requests. Luckily, we had plenty of humans on hand to save the day and to train our NLP systems. We iterated almost every hour to make our AI smarter.

In the end, the event served as a useful stress test, helping us identify new areas for future strategy and improvement. It also served as an important lesson to start small and scale over time.

When scaling AI, don’t go and code in your entire interaction from the get-go. Instead, bootstrap it with humans and then learn to master the hand-off. This approach gives your AI an important scaffolding for getting itself off the ground and creating a codified system to iterate on quickly. Then, as the AI scales and gets smarter, simply remove one set of training wheels and trade them in for the next.

Snehal Shinde is the CTO and cofounder of Mezi, the travel and shopping app.